Machine learning in trading: theory, models, practice and algo-trading - page 3185

New comment
 
fxsaber #:

There are several hypotheses.

I compared the characteristic "maximum potential profit". I did not see any significant differences.

 
fxsaber #:

I compared the characteristic "maximum potential profit". I did not see any significant differences.

Is that if you do it in a ticky-tack way? And if you're trend-following? Compare from reversal to reversal.
 
Forester #:

Amount1 - Amount2 is more like volatility. Trendiness is if you add up a lot of them. In real data, trends are one, in randomised data (up to about through1), trends are more like random outliers due to increased volatility. I assume that they are many times smaller in amplitude than the real ones.

UPD: I didn't see that you have ~ instead of - there.

About ~. Their approximate rave means just that, real well mixed, averaged through 1.

I look at a trend as a shift in the mean in increments relative to zero.

but it's all a matter of taste, I guess.

par(mar=c(2,2,2,2),mfrow=c(2,1))

mn_trend <- c(rep(-0.5,100),rep(0.5,100))
rn <- rnorm(200)
cbind(rn , mn_trend) |> matplot(t="l", lty=1, col=c(8,2),main="random and mean")

rn_trend <- rn + mn_trend
rn_trend |> cumsum() |> plot(t="l",main = "cumulative sum rn + mn_trend")
 

Forester #:
Это если по-тиково?

Yes.

And if you're trend-following? From reversal to reversal compare.

Started to change the size (within reason for scalping) min. knee ZZ and watch the sum of the knees.

The random symbol has such potential profit higher than the original symbol. That is, the random symbol is potentially more profitable.

If the potential profit was lower than the original, it could somehow explain the failure with scalping. But here it is the opposite situation.


ZЫ In general, if there is an interest to try to find differences between the two series, they can provide them.

 
fxsaber #:

Yes.

Well then you should know that in spectral analysis for example with a hundred harmonics you can describe a series of 10,000 values with pretty good accuracy....

take 10,000 values ---> get ---+ the same thing but 100 values.

It is absurd that 100 values can describe the original series of millions of values! It seems to be a tool of theorists but not of practitioners.

And you call it absurd, weird....

fxsaber #:

Unfortunately, it is not clear what input parameters of randomisation to optimise.

Everything I write here is purely my fantasies, so look critically....


You can try to create from the same harmonics a series on which your TS will work ...

The parameters for the optimiser are a combination of harmonics,

The fitness function is the quality of the TC's performance on this synthetic data.

 
fxsaber #:

A random symbol has such a higher potential profit than the original symbol.

That's weird.

So it passes the Monte Carlo test. If there is profit on the real ones and not on the jumbled ones.

 
mytarmailS #:

Then you should know that in spectral analysis, for example, a hundred harmonics can describe a series of 10,000 values with pretty good accuracy...

take 10,000 values ---> get -+ the same thing but 100 values.

And you call it absurd, that's strange...

Because it has nothing to do with the cvr. mp3 and jpg even at very low bitrates are recognisable by neuron. But alpha in the form of scalping is lost even if the "bitrate" is maintained.

 
fxsaber #:

Because it has nothing to do with cvr. mp3 and jpg are recognisable by neuron even at very low bitrate. But alpha in the form of scalping is lost even if the "bitrate" is preserved.

It's all numbers and conversions what do you mean it has nothing to do with cvr ? it's just numbers....

It's like saying that a picture of dogs has nothing to do with a picture of kitties... because kitties are not dogs...

And what is bitrate?

 
Forester #:

So it passes the Monte Carlo test. If the real ones are profitable and the mixed ones are not.

If this test is considered as a clear sign of difference between real and randomised series - yes, 100% passes.

My "monte carte" is to create a lot of scalping histories. And on them to identify the vulnerabilities of TC. Right now there is stupidly not enough history length for such checks. That's why we need adequate generation.


The idea of generation seemed even beautiful, I have not seen anything like it. But it turned out that it is not suitable for my purposes.

But the Monte Carlo test, indeed, passes with flying colours. But this is a side effect, which is of little importance.

 
mytarmailS #:

It's all numbers and conversions what do you mean it has nothing to do with the WRC ? it's just numbers....

Answer yourself what useful characteristic is being retained/lost. I answered that question with my conversion.

And what is bitrate?

Width of information flow (in network context).

1...317831793180318131823183318431853186318731883189319031913192...3574
New comment